0:00

Okay, this year has been called the

0:02

year of AI agents. Now there's a

0:04

ton of buzz about agents on YouTube,

0:06

on X, all of these different platforms,

0:08

and actually a lot of it is

0:10

wrong. A lot of these agents are

0:12

not very good yet. We want to

0:14

give you the actual real download of

0:16

what is happening with agents and we're

0:18

doing that with one of the best

0:20

brains in the business, Joe Mora, the

0:22

CEO and founder of Crew, AI. We're

0:24

actually going to break down what is

0:26

an agent. We're going to tell you

0:28

where you can get started to build

0:30

agents for your role today and where

0:32

they'll have impact. We're going to give

0:34

you real use cases, things that agents

0:36

are actually doing and make an impact

0:39

for businesses today in marketing and sales

0:41

in other places. And stay tuned because

0:43

we're going to tell you your job

0:45

in the future is going to depend

0:47

on the quality of the agents you

0:50

have built to help you do your

0:52

role. All that and more on this

0:54

episode of Market Against the Grain. I'm

0:56

your co-host as always Kieran Flanagan. Here

0:58

as always for my co-host Kip Bodner.

1:00

Let's get into the day show. Here's

1:03

a quick message from Hubspot.

1:05

This isn't your typical marketing

1:07

software at. Because Hubspot isn't

1:09

typical. It's marketing made easy.

1:11

Easy. Easy. Turn one piece of content

1:13

into assets for every channel. convert

1:16

leads in no time and get

1:18

a crystal clear view of your

1:20

campaign performance. Hubspot can do all

1:22

that and get your results fast,

1:24

like double your leads in 12 months

1:27

fast. See, I told you

1:29

this wasn't a typical software

1:31

ad. Visit hubspot.com/marketers to get

1:34

started for free. Okay,

1:47

we're here with Joe Mora the CEO founder

1:49

of crew AI crew AI is your

1:51

global control pain for agents One of

1:53

the best minds on everything there is

1:55

around agents Joe very happy that you're

1:57

joining the show Hey there, thank you

1:59

so much for having you. So Joe,

2:01

this is like the year of agents,

2:03

right? This is all we've heard. I

2:05

actually have been paying attention to a

2:07

lot of the things happened this year.

2:09

I think it was World Economic Forum

2:11

and there was another big meetup of

2:13

all the AI minds. And I think

2:15

the number one thing on everyone's lips

2:17

is like. agents agents agents agents agents

2:19

right that's most of what we are

2:21

all talking about and so I thought

2:23

we could just tee up for our

2:25

audience maybe we'll actually start with what

2:27

is an agent maybe you can explain

2:29

to our audience like what do we

2:31

mean by agent for one of the

2:33

best minds here to explain agents and

2:35

we can actually get into your take

2:37

on Is it really the year of

2:39

agents? And what do we even mean

2:41

by that? Like where do you really

2:43

see agents be an impactful this year?

2:45

So maybe start with like what is

2:47

an agent? Like maybe explain to our

2:50

audience how you think of it. Well,

2:52

first of all, thank you so much

2:54

for having me. I'm so excited that

2:56

we get to talk about this and

2:58

yes, everyone was to talk agents and

3:00

I gotta say that that has been

3:02

an interesting year so far. But that's

3:04

a great way to start, right. So

3:06

the way that we think about it

3:08

is everyone knows about all these LLLM,

3:10

so Judge EPT and Tropic and everything.

3:12

So they're very good at kind of

3:14

like predicting content, right? So if you

3:16

really say like, hey, write me an

3:18

email, it will do it for you.

3:20

And if you say, well, make it

3:22

funnier, it will do that for you.

3:24

Now, the interesting thing is it almost

3:26

has some sort of cognition, right? If

3:28

you give them two options of emails

3:30

of emails, it's gonna like choose between

3:32

the two and give you're reason for

3:34

that. So the beauty of agents is

3:36

you can exploit that feature to how

3:38

this LLLM's kind of navigate a problem

3:40

on their own. So it's not a

3:42

chat anymore. You give it a task

3:44

and you can leave the room and

3:46

then this agent's gonna try to autonomously

3:48

kind of like through this idea of

3:50

reasoning figure out how to get there.

3:52

So I would say that the definition

3:54

of an agent is you've got to

3:56

have agency. Yeah, and it's got to

3:58

have like some components, right? Like it

4:00

has tools, it has memory, like maybe

4:02

just talk about some of the common

4:04

characteristics of what an agent will have.

4:06

Yeah, because if you think about the

4:08

LLLM on the silo, it just speed

4:10

test out, right? But in order for

4:12

you to make this more agentic, you

4:14

need to have a way to hold

4:16

that information. So you're going to need

4:18

like some sort of memory. And then

4:20

there are going to people that are

4:22

going to ask about shorter memory and

4:24

long-term memory. dive into that and things

4:26

get very technical. But there's memories and

4:28

there's also tools. I would say those

4:30

are the two big components. So as

4:32

the agents are trying to do something

4:34

they're going to use those tools to

4:36

interact with other systems. ERP or CRM,

4:38

whatever that might be. Right. Really when

4:40

you think about an agent, most of

4:42

them will have memory because you need

4:44

that. to have, as you said, agenda

4:46

behavior, most of them will have access

4:48

to tools because they can actually do

4:50

things on your behalf. They're autonomous. And

4:53

so as we sit here today, one

4:55

of the things that Kip and I

4:57

were Jammonon earlier on, and when I

4:59

say Jammonon, I mean, Kip showing me

5:01

because I'm a European and I live

5:03

in the dark ages and we are

5:05

not allowed to have access to anything

5:07

without the bureaucrats signing it off for

5:09

us. And so I do not have

5:11

access to... ChatGBT operator today, other than

5:13

probably through a VPN at the weekend.

5:15

But it's a good segue into like

5:17

the year of the agents and how

5:19

you think about that. So Open AI

5:21

launched that. I think that Kip, your

5:23

take was, it's cool for geeks like

5:25

us. Why don't you give us your

5:27

take, Kip, I don't know where's in

5:29

your mind, and then I would love

5:31

to get Joe, your take on their

5:33

launch and just where we are in

5:35

agents in general in terms of capabilities

5:37

capabilities. In terms of capabilities. as part

5:39

of their chat gPT pro edition, which

5:41

is the $200 month edition. So I

5:43

immediately had to upgrade from $20 a

5:45

month to $200 a month. They took

5:47

your money. They took my money like

5:49

quick. I want to know one pro

5:51

and everything too. So I got to

5:53

do that upgrade. So there's already a

5:55

cost barrier. And what it does. is

5:57

it has basically a browser control agent

5:59

where opening eyes built their own browser

6:01

and the AI can go and navigate

6:03

based on a request. So you could

6:05

make hotel reservations, flight reservations, dinner reservations,

6:07

do research and it's kind of slow.

6:09

It's kind of clunky. If it's a

6:11

good background task and it doesn't require

6:13

me to like log in or do

6:15

a bunch of interaction, it's just like,

6:17

hey, can you look up a hotel

6:19

with a pool in the pool in

6:21

the city? That's great it can do

6:23

that real quick and give me all

6:25

the information real slow it doesn't do

6:27

real quick it doesn't much lower than

6:29

like I know it's like 10 times

6:31

slower than a human doing it right

6:33

yes you could have went to upwork

6:35

hired a freelancer to be your personal

6:37

searcher and had them search for it

6:39

in the time that open AIs to

6:41

operate or actually completed the task look

6:43

it is very slow right now I

6:45

think it's a peek at what's to

6:47

come right I'm interested Joe and your

6:49

take but it's a look at the

6:51

future is a little ways away Well,

6:53

I'm going to say, I agree with

6:56

you, but I have, I think, would

6:58

be a hot take. Maybe not a

7:00

hot take. I don't know. And that's

7:02

how I started to boot crew, right?

7:04

My first experience with crew was building

7:06

agents that would help me with like

7:08

posting things on socials. I was never

7:10

really good at that. I had all

7:12

these ideas, but from getting that idea

7:14

and making that into a well draft.

7:16

thing that if you feel comfortable putting

7:18

out there in the word like there's

7:20

a huge gap in there right and

7:22

I could do it for sure it's

7:24

just a matter of that you're gonna

7:26

have to sit down you're gonna have

7:28

to spend one hour kind of like

7:30

doing this and if you want to

7:32

do it consistently you're gonna have to

7:34

do it every day and I was

7:36

never able to do it the minute

7:38

that I got agents to help me

7:40

with that I start doing it every

7:42

day because we're right they would take

7:44

way longer but now I could speed

7:46

out my crazy idea and I would

7:48

make a coffee, I would get to

7:50

work something, I would forget that the

7:52

thing was running, and then I would

7:54

go back and it was ready. I

7:56

was like, all right, this is good.

7:58

So I hear you, I think. they're

8:00

not optimal or faster than humans by

8:02

large. There are ways that you can

8:04

get there, specifically if you find doing

8:06

certain models for certain things, but I

8:08

do see a lot of value, especially in kind of

8:10

like this idea of your firing a bunch

8:12

of them in the background, and you're firing

8:15

a bunch of them in the background, and

8:17

you've got to do something else and just

8:19

not think about it. As you said, like,

8:21

I agree. What Kip said as well is

8:23

like, the latency doesn't matter if it's a

8:25

task that it's a task that you just

8:27

want to So here's a good example of

8:29

something I think operator would be really good

8:31

at. Let's say you're moving into a

8:33

new apartment, you have a room,

8:36

and like maybe you've picked out a bed

8:38

or something. You can upload the picture of

8:40

the room, you can upload a picture of

8:42

the room, you can upload a picture of

8:45

the bed, and you can upload a picture

8:47

of the bed, and you can upload a

8:49

picture of the bed, and you can have

8:51

those things actually go together, and it just

8:54

happens in the background, and that stuff that

8:56

you would have Right. The long-term goal to

8:58

have a personal assistant, which is what they're all

9:00

building towards. There's one other thing I want to

9:02

just touch on here because it's important of agents

9:04

in general, Joe, before we kind of get into

9:07

just your take on what you think about.

9:09

the impact agents will make this year.

9:11

And it actually is the UX pattern that

9:13

Open AI had in the operator, because one

9:15

of the big question marks around autonomous agents

9:17

is, what's the right UX patterns so people

9:20

feel comfortable with them? And what I mean

9:22

by that is, you know, these agents are

9:24

autonomous, so when KIP asked it to book

9:26

a table on open table, should it just

9:29

come back and say, it's booked, or should you

9:31

actually be able to like see the agent

9:33

complete the task, so you feel... comfortable with

9:35

the agent doing something on your behalf and

9:37

they went down the path to have a

9:39

UX pattern where you can see the agent

9:41

doing its work and at any point in

9:43

time you can take control away from the

9:45

agent. But what's your general take on autonomous

9:47

agents in terms of how comfortable humans are

9:49

going to be to integrate them into the

9:51

workflow like the right kind of UX pattern?

9:53

Yeah, that's a great question. I actually spent

9:55

some time talking with some argument about it

9:57

and it was pretty good because if you

10:00

it's like AI is moving very fast,

10:02

right? So it's not waiting for its

10:04

native protocols. Because if you think about

10:06

it, it should not even use a

10:08

browser, right? A browser doesn't make sense.

10:10

The concept of, can like the buttons

10:12

and everything, that doesn't make any sense,

10:14

or that doesn't make any sense, or

10:16

even keyboard and mouse, that just increases

10:18

latency and reduce throughput for these models.

10:20

Like, in matter of fact, they should

10:22

not even use language to communicate to

10:24

communicate with language to communicate with what

10:26

is happening there. There's a security aspect

10:28

of it because you can always see

10:30

what is happening right now. I just

10:32

think it's getting better and better. You

10:34

want to make sure that you're able

10:36

to do that. But there's also that

10:38

ability for you to feel reassured about

10:40

it. So I think it's going to

10:42

come to a time where you're going

10:44

to feel good about just... find requests

10:46

into the void and things are going

10:48

to get done for you. But I

10:50

think right now people just wanted to

10:52

feel they are more in control of

10:54

these things, right? And this is true

10:56

not only on the personal level, but

10:58

also on companies. If companies are thinking

11:00

about, and these were seeing first-handed on

11:02

our enterprise deals, like if a company

11:04

was to automate a critical part of

11:06

their process, they want to make sure

11:08

that they can visualize and control this

11:10

and they understand what is happening, and

11:12

they can out it later on. So

11:14

I think it's a temporary... that might

11:16

change in the future, but I just

11:18

don't know yet. Yeah, I was talking

11:20

to someone who had been using very

11:22

early on when agents first come out,

11:24

they are playing with all the technology,

11:26

and they were set up an autonomous

11:28

agent to do some stuff on email,

11:30

and the agent was meant to craft

11:32

emails, put them in the draft, and

11:34

then they would go in and look

11:36

at them and then send them or

11:38

not send them. and they had realized

11:40

that the agent started sending them. They

11:42

were sending some like pretty bizarre emails.

11:44

They were like, yeah, like I feel

11:46

the great UX pattern is I need

11:48

to be able to see what the

11:50

agent's doing. And so maybe that's a

11:52

good segue to like, you know, there's

11:54

a lot of, hey, this is transformational

11:56

this year. Agents are going to be

11:58

part of the workforce that was the

12:00

common talk thread over the past couple

12:02

of weeks from every tech leader there

12:04

is. What do they mean by that?

12:06

at what's happening with agents, what do

12:08

you think they mean by that and

12:10

what are you bullish on and what

12:12

are you not bullish on when it

12:14

comes to agents? Yeah, that's a good

12:16

one. Well, first of all, I think

12:18

I want to show you a visual

12:20

real quick because that will help me

12:22

to make my point and that is

12:24

agents are happening, right? So in here

12:27

what you're seeing is the number of

12:29

cruise executions per month. This was pulled

12:31

a few days ago, so January is

12:33

still going. and up to this point

12:35

has been over 16 million crews. Each

12:37

crew has many agents within it. Some

12:39

have up to 21, that's the highest

12:41

number that I have seen, but you

12:43

can go higher than that. So what

12:45

we're talking here is tens of millions

12:47

of agents being executed every month. And

12:49

the reason why I want to double

12:51

click on that is just that it

12:53

is happening. In my mind, it is

12:55

happening the genie is not getting back

12:57

into the bottom. Now it's a matter

12:59

of how fast it will happen. and

13:01

how good you get in what period

13:03

of time. But I think what a

13:05

lot of what we're seeing on companies

13:07

being bullish about it is because, and

13:09

I'm going to take one step back,

13:11

a lot of people are comparing kind

13:13

of like agents with kind of like

13:15

the internet early days, and I think

13:17

that's an interesting way to correlate the

13:19

both. But if you were online on

13:21

the internet on day zero, you would

13:23

have no upside, no impact on your

13:25

bottom line, because no one was online.

13:27

But what we're seeing with AI is

13:29

companies implementing it in like two quarters

13:31

later They're reporting impacts on their bottom

13:33

line. So you've got some of those

13:35

10 K's and 10 Q's and you

13:37

see companies like Walmart Like starting to

13:39

save millions of support You know like

13:41

all right, something is happening here. So

13:43

I think there's a mandate on the

13:45

executive level on this company is on

13:47

the board level of this company is

13:49

that this is happening three years from

13:51

now we need to figure it out

13:53

how we're going to manage deploy these

13:55

resources and from the edge side again

13:57

everyone's tinkering with it and decided so

13:59

it's almost like a claw motion where

14:01

like incentives are aligning and I think

14:03

that is just fooling this up. So

14:05

I'm bullish that this is the year

14:07

of agents and what I mean by

14:09

that is people are going to deploy

14:11

a lot of agents this year. They're

14:13

going to try a lot of things

14:15

this year. Now is this the year

14:17

where companies are going to automate entire

14:19

departments? I don't think just yet, but

14:21

I think this year is where things

14:23

starts to get very pretty seriously.

14:26

Could you maybe tell us? how companies that

14:28

are deploying these crews and agents

14:30

are doing it in the right

14:32

way. So to your point, the

14:34

wrong way is probably I'm going

14:36

to go in and create agents

14:38

to replicate what these humans are

14:40

doing, versus a lot of the

14:43

success I have seen in agents

14:45

is like, they complete micro tasks. that

14:47

make up part of your role freeing

14:49

you up to actually spend time on

14:51

things that are much more important right

14:54

so that actually the human can become

14:56

much better at that role but could

14:58

you maybe talk to us like where

15:00

do you see companies deploying them in

15:02

the right way and like what examples

15:04

kind of use cases are you seeing that

15:06

work really well? Yeah so I actually

15:09

spent a good time talking with Jacob

15:11

Wilson he's the commercial CTO at WW

15:13

and Gen AI. And it's amazing to

15:15

work with them. They're using crew a

15:17

lot. And one big thing that we

15:19

talked about, and there's a whole interview,

15:21

I can send a link over to

15:23

you if people want to watch, is

15:25

there is a cultural aspect on adopting

15:28

agents in the company, right? People like,

15:30

people fear that, or people trying to

15:32

understand what role do they play. The

15:34

companies they're being most successful, and PWC

15:36

is one of them. What they are

15:38

doing is they're kind of promoting people,

15:40

right. So, you're still accountable for the

15:42

end of Brazil, you're still accountable

15:44

for reviewing this, but a nice

15:46

ball on it, present this, but

15:49

now you have this extra tool

15:51

and these agents can do something for

15:53

you. So, a cool example that I

15:55

can mention is we're working with a

15:57

telecom company on a legal use case.

16:00

like they have their legal like

16:02

people that can do everything but

16:04

what is happening now is they're

16:06

ultimately a lot of the contract

16:08

analysis with these agents beforehand. So

16:10

by the time that it gets

16:12

to legal, it already has recommendations,

16:14

red lines, a bunch of other

16:16

things. So kind of like basically

16:18

scaling things that they couldn't do

16:20

before. We thought it being prohibitive

16:22

expensive. But there's so many more

16:24

use cases. And we can talk

16:26

about the sales and the marketing

16:28

use cases and the back office

16:30

automation. There's quite a lot going

16:32

on out there. we've talked a

16:34

lot about kind of where agents

16:36

are and now I was trying

16:38

to understand like what are the

16:40

core use cases that people are

16:42

actually building? What should people go

16:44

and do? There are a lot

16:46

of people who watch our show

16:48

who are like, hey I just

16:51

want to like understand what I

16:53

should be doing, how I go

16:55

out and build an MVP of

16:57

that to see if it actually

16:59

makes sense for me, my business,

17:01

what have you? Like what's happening

17:03

and where should people start? These

17:05

are the most common horizontals within

17:07

a company that we're seeing agents

17:09

and use cases being deployed. Now,

17:11

there's a few interesting things that

17:13

you can infer from this. One

17:15

is there is no clear winner,

17:17

right? There's no like, oh, people

17:19

are using for marketing alone. No,

17:21

it's very much spread out. What

17:23

for companies like us is good

17:25

news, because it means that you

17:27

can land and expand into other

17:29

areas, right? But if you interview

17:31

these people as we did and

17:33

we talk closer with them, the

17:35

common pattern is actually starting with

17:37

simpler use cases. It's what we

17:40

call low precision versus high precision.

17:42

So low precision use cases, they

17:44

require, let's say, 9% certainty or

17:46

accuracy on their outputs, but high

17:48

precision use case required 99.99. So

17:50

an example of a low precision

17:52

could be, well, I want agents

17:54

that will help me draft presentations

17:56

for sales calls out of cross-transcripts

17:58

or my CRM information. and high

18:00

precision use cases that we are

18:02

seeing out there is helping companies

18:04

and banks fuel IRS forms, where

18:06

like you don't want to get

18:08

that wrong, right? Yeah, you can't

18:10

mess up your tax forms. It's

18:12

like you're talking about a big

18:14

corporation. Just an anecdote, funny enough.

18:16

Some of these forms, they are

18:18

70 plus pages long, but they

18:20

call me for instruction manual. That

18:22

is 620 pages long. So yeah,

18:24

agents can help with that, but

18:26

that's more high precision use cases,

18:29

right? You're gonna, those start there,

18:31

you might get a burn. Start

18:33

with low precision and scale from

18:35

there. That's really great advice, I

18:37

just want to kind of recap

18:39

that for the audience. So if

18:41

you were like in a role

18:43

and you wanted to even start

18:45

with agents before you've even started

18:47

to get into how to build

18:49

them and we can get into

18:51

that and show you some use

18:53

cases, what your recommendation is is

18:55

like look. And AI can help

18:57

with this because I've actually used

18:59

it as a test for this.

19:01

If you actually take a role,

19:03

let's say your rule is a

19:05

BDR and you split out your

19:07

tasks into low precision, high precision,

19:09

and you can actually start to

19:11

pick some ones that are in

19:13

that low precision category, and that's

19:15

like some places to experiment with

19:18

agents. Is that like the right

19:20

way for someone to get started?

19:22

Yeah, the way that we put

19:24

it is basically you have four

19:26

bullet points. What is... Be another

19:28

adopter, you want to get ahead.

19:30

Two is don't wait for other

19:32

people use cases, right? A lot

19:34

of people want to know what

19:36

are like what is company acts

19:38

doing? Like you don't want to

19:40

go there. So start simple. That's

19:42

basically what you're saying. Start with

19:44

something simple. And when you expand,

19:46

you want to expand into kind

19:48

of like the low risk, high

19:50

impact kind of like use cases

19:52

and go in that direction. But

19:54

yeah, exactly as you said. And

19:56

then in order for you to

19:58

view them. there's so many different

20:00

ways right now, right? Like there's,

20:02

and if anything, there's gonna be

20:04

even more. There's, for less technical

20:07

people, there's. like no code platforms.

20:09

And we offer that as well,

20:11

for example, Korea AI, but there's

20:13

many others. Like if you're a

20:15

more technical person, you can use

20:17

frameworks like Korea itself, where you

20:19

can actually code in Python, some

20:21

of this agents, and some of

20:23

this prompt. So I think there's

20:25

a bunch of different flavors that

20:27

people can use. Now, a lot

20:29

of the true value gets unlocked.

20:31

When you get more technical people

20:33

involved, I would say you got

20:35

that right. And that is it

20:37

starts with kind of like low

20:39

precision use cases and simple and

20:41

then expand into low risk high

20:43

impact. Right. Let me tell you

20:45

about a great podcast. It's called

20:47

Creators of Brands. It's hosted by

20:49

Tom Boyd. It's brought to you

20:51

by the Hubspot podcast network. Creators

20:53

are brands, explores how storytellers are

20:56

building brands online. From the mindsets

20:58

to the tactics to the business

21:00

side, they break down what's working

21:02

so you can apply that to

21:04

your own goals. Tom just did

21:06

a great episode about social media

21:08

growth called 3K to 45K on

21:10

Instagram in one year, selling digital

21:12

products and quitting his job to

21:14

go full-time creator with Gannon Mayer.

21:16

Listen to creators or brands wherever

21:18

you get your podcast. So maybe

21:20

we want to get in to like

21:22

show some of these use cases, but

21:24

I do think an important point to

21:27

make is, okay, I've decided that I

21:29

understand what an agent is, I've looked

21:31

at my role and I understand like

21:33

what are good use cases for an

21:35

agent. The other thing you mentioned is

21:38

like how agents are being adopted within

21:40

the company. like how they're being built

21:42

for different teams. Can you maybe just

21:44

talk a little bit about that trend,

21:46

like how you see companies adopting agents?

21:49

Yeah. And you mentioned it's like still

21:51

quite a technical thing and maybe kind

21:53

of just touch on that. Yeah. for

21:55

sure. I think it's interesting because early

21:57

days what he would have with like

22:00

LLLMs and I think a lot of

22:02

people probably experiences this first-handed is you

22:04

would have on the edges of the

22:06

organization on individual teams people would just

22:08

pick up LLLMs to do work for

22:11

them. So someone that's a little more

22:13

savvy would start to play around with

22:15

chat cheap and they would found this

22:17

amazing prumps that would help them with

22:19

kind of like customer proposals right. And

22:22

then that became a problem because Well,

22:24

this person is putting information that it

22:26

might not be supposed to be there,

22:28

the company doesn't approve, and maybe this

22:30

person unlocked an amazing case, but that

22:33

information now and that knowledge is silo.

22:35

So while we are seeing with agents,

22:37

at least on the enterprise sales motion

22:39

for Korea, it's more of a central

22:41

deployment. So usually under a CIO, a

22:44

CTO, a head of AI, and you

22:46

sell into this department that is going

22:48

to configure everything, and then start to

22:50

enable these departments individually. And the cool

22:52

thing about that is you're going to

22:55

have way more control on what are

22:57

the elements that are being used. Do

22:59

I want to add future short PII

23:01

and personal information? And making sure that

23:03

I'm controlling all these use cases so

23:06

they are reusable. and then even enabling

23:08

people that can code to use the

23:10

platform to build it. So that's kind

23:12

of like what we're seeing in terms

23:14

of the enterprise adoption. And again, funny

23:17

enough, like there's a lot of kind

23:19

of like low code, easy to roll

23:21

out and templates that you can use.

23:23

But once that you want to get

23:25

into those very kind of like chunky

23:28

use cases, right? Because if you forget

23:30

the name AI agents for a second,

23:32

we are talking about AI powered customized

23:34

process because all these companies have no

23:36

equal process. It's when technical people and

23:39

be involved unlocks a lot of value,

23:41

especially on integrations with homegrown systems inside.

23:43

Right. Yeah, like they're much more powerful

23:45

that they're deeply intertwined within your unstructured

23:47

data, and they're deeply intertwined within your

23:50

systems. Yeah, but I gotta say, it's

23:52

still very much early day, so we

23:54

basically pulled around 4,500 people from different

23:56

companies, and only 15% actually have features

23:58

in production right now. If you isolate

24:01

this, only looking at large enterprises, then

24:03

we're talking about 23, so it's a

24:05

higher number. What can the hints that

24:07

enterprises are moving a little faster here?

24:09

But for a lot of people, it's

24:12

in like early days or... They just

24:14

have done a few deep dives. So

24:16

I think this is very interesting to

24:18

watch. What is that Joe? The 15%

24:21

is enterprise companies that have used an

24:23

agent to complete a task? Or what's

24:25

the 15%? No, that a 15% is

24:27

overall companies, not only enterprises, that have

24:29

features with AI agents in production. And

24:32

if you cut off and look at

24:34

only the enterprise, that number bumps into

24:36

23%. Okay. But that is having like

24:38

AI agent powered features in production from

24:40

the companies that we're talking with. Not

24:43

necessarily in their product, it might be

24:45

like internal automations, right? Might be a

24:47

backoffs automation. Okay, so it is internal,

24:49

yeah. I think internal is actually, this

24:51

is one thing that I found very

24:54

funny. I mean, I had this hypothesis

24:56

in 2024 that I would see way

24:58

more SAS companies kind of like adopting

25:00

agents, just kind of like trying to

25:02

reinvent themselves. But funny enough, I'm not

25:05

seeing a lot of that. It's a

25:07

lot of more traditional organizations that are

25:09

kind of like trying to figure out

25:11

how to get more efficient. In the

25:13

SAS companies that we have talked with,

25:16

I get more reluctance from them on

25:18

it. I'm not sure what that it's

25:20

coming from. that have been proving wrong.

25:22

I feel like that number is higher

25:24

than I would have expected. I would

25:27

have expected like single-digit adoption of like

25:29

agents being deployed. It just it shows

25:31

that even though it's early there is

25:33

a lot of just crappy stuff that

25:35

has to get done out in the

25:38

world that agents are good enough to.

25:40

go and solve today, right? Or otherwise

25:42

that number would be much, much lower

25:44

than it is in terms of like

25:46

agents deployed. What I would say though

25:49

is there is one caveat, right? These

25:51

people, they were coming into Korea to

25:53

answer this form, so they are definitely

25:55

more savvy. They're looking into agents, so

25:57

they're definitely more savvy. They're looking into

26:00

agents, so I think if you look

26:02

at the broader population of like any

26:04

company out to like the... Lagarts. We're

26:06

still so early. I think the point

26:08

here is like, if you're listening to

26:11

the podcast and you're following along and

26:13

you're thinking, well, I'm going to create

26:15

an agent and do something, you're in

26:17

the fast movers. Yes. Exactly. I think

26:19

one of the things that would really

26:22

help our audiences showcase some of the

26:24

agents that have been built for real

26:26

use cases and go through like what

26:28

these agents are actually doing. for maybe

26:30

a salesperson or a marketer or something

26:33

that you really think is like a

26:35

good example of an agent in action.

26:37

Yes, so I think on the marketing

26:39

and sale side, one of the most

26:41

interest ones that I have seen is

26:44

agents that are doing a couple facts.

26:46

They start with enrichment. So the way

26:48

that they deploy this was actually an

26:50

in-product feature and a back office, kind

26:52

of like automation. So when someone would

26:55

come into their website and create an

26:57

account, they would have agents go online

26:59

and start researching this person. So up

27:01

to this point, that's okay, right? This

27:03

is kind of like regular enrichment, like

27:06

find what is this person role, find

27:08

like more about this company vertical and

27:10

all that. So that's good. Now, where

27:12

things starts to get interesting is they

27:14

got some agents to get one step

27:17

ahead and given the information that they

27:19

found come up with hypothesis on how

27:21

this person is going to use their

27:23

product. So what is the interest? If

27:25

this person is kind of like a

27:28

CMO at a company, how they are

27:30

going to use like what value they're

27:32

going to get from this product and

27:34

creates kind of like three hypothesis. And

27:36

then do the same thing for the

27:39

company, like for a company in this

27:41

vertical, what do you believe in like

27:43

the main three things that would take.

27:45

So now that the agents have done

27:47

that, they convert it into Jason, so

27:50

structure data, and then push it in

27:52

two places to their hub spot and

27:54

into their product database. So what that

27:56

means is now out there a meal

27:58

marketing. It's like. super hyper target, mentioning

28:01

not only the name of the person,

28:03

the company, but like highlighting the features

28:05

in the ways that they could leverage.

28:07

And then in the product, and this

28:09

kind of like what they're working on,

28:12

in the product, they preview some of

28:14

the information and the templates that they

28:16

show, like basically using that inference that

28:18

they made. So very interesting use case

28:20

case, and I think a great kind

28:23

of use case for agent in general.

28:25

Yeah, it's somewhat similar to what we

28:27

do. And I think the example going

28:29

back all the way to what an

28:31

agent is and having memory tools, like

28:34

an example of a tool that is

28:36

really good for research is something like

28:38

perplexity's new sonar, like perplexity is a

28:40

pretty good research tool that the agent

28:42

can have access to and do some

28:45

of that research on your behalf. So

28:47

I think that's a good example of

28:49

like the agent being able to autonomously

28:51

complete those tasks without any real human

28:53

in the loop and you as a

28:56

sales rep have that stuff. at hand

28:58

and so you can be much more

29:00

productive because you have this stuff being

29:02

done for you by a small team

29:04

of agents who each are like specialized

29:07

in one of those tasks. Exactly and

29:09

I think we go back to the

29:11

comparison that we make with the operator

29:13

early on right like yes you as

29:15

a sales rep you could go out

29:18

and research this customer and make sure

29:20

that you're engaging them like on a

29:22

very custom format or like prepare for

29:24

every meeting that you have yes you

29:26

absolutely you absolutely you can. You might

29:29

be able to do it faster than

29:31

what an agent would do, but if

29:33

you can do that at scale, that

29:35

means that you can take more meetings

29:37

and be better prepared to them. And

29:40

there's also something beauty about you being

29:42

able to customize it, right? Let's say

29:44

that you always have that when say

29:46

it was rap in your company, then

29:48

it's just like the beast, right? The

29:51

guy does the best rap. Why don't

29:53

you just like try to replicate that

29:55

for everyone and now everyone has that

29:57

level of notes? I mean, the sales

29:59

rap plan would be too happy about

30:02

it, but it might be a definitely

30:04

an interesting experiment. I would love to

30:06

get into another use case you mentioned

30:08

off, Mike, but I think one of

30:10

the interesting things here, because it was

30:13

a point you made which is... the

30:15

rep can have these agents and be

30:17

much better at their role and other

30:19

reps might not be happy about it.

30:22

I could create some competition to actually

30:24

use agents because you can't compete unless

30:26

you have access to the same sort

30:28

of help. And I don't know if

30:30

you saw there was an interview from

30:33

Satya Nadala, the Microsoft CEO this week,

30:35

and one of the things he said

30:37

was in the future, people might be

30:39

hired. because of the agents that they

30:41

have helping them do their role. And

30:44

so he can imagine a world where

30:46

you go to LinkedIn and instead of

30:48

having certifications or anything like that, and

30:50

even prior work experience, you actually list

30:52

at the agents that you have built

30:55

to help you be like the kind

30:57

of employee that you are. Let's just

30:59

get your kind of thoughts on that

31:01

as a future version of what like

31:03

a great employee might look like. I

31:06

a thousand percent agree. Honestly, not because

31:08

of the agent themselves, but also because

31:10

why do I tell about that person?

31:12

And I can tell that we are

31:14

doing this at Korea. So for example,

31:17

when we interview people for engineering roles,

31:19

we tell them during the interview, you

31:21

are allowed to use anything. Like you

31:23

can use chat cheapity, you can use

31:25

entropic, you can use coarser, you can

31:28

search Google, whatever it is, you're allowed

31:30

to use. If they don't use it,

31:32

It's an automatic pass. And how well

31:34

they use it actually comes a lot

31:36

on if we're going to make them

31:39

an offer or not. And that can

31:41

be counterintuitive for a lot of people

31:43

like, well, but are you assessing like

31:45

their engineering skills? Well, on the day

31:47

to day, they're going to have access

31:50

to these tools anyway. So I want

31:52

to know how well they use it.

31:54

And I would say like probably what

31:56

he's hinting there is some of that

31:58

as well. the agent that they have

32:01

as their companion, but also the

32:03

ability they were able to create something like

32:05

that and might tell about that person.

32:07

Yeah. Coming back to the use cases you

32:09

mentioned off, Mike, I think one of the

32:12

ones we should actually just cover real quick

32:14

because everyone loves a good SEO content use

32:16

case, could you maybe just go through that

32:18

use case to just give another example

32:20

of what agents can help with? Yes.

32:22

I love that one. That was so

32:25

good. So this was a stockup that

32:27

we were helping on using agents. It

32:29

was very early days in crew. It

32:31

was very interesting. And what they wanted

32:33

to do is everyone likes to do

32:36

not only SEO, but overall conversion

32:38

rates, right? And AB testing is such

32:40

a big thing. And everyone knows that

32:42

it works, but takes a lot of

32:44

work to do it well. So what

32:46

we were building together was agents that

32:49

would like give an ER. product goes

32:51

into our website, would take screenshots

32:53

of our website, understand like the

32:56

copies, everything. Then when the research, what

32:58

are their competitors, would go into their

33:00

websites and look at all their copies

33:02

and everything, and then we'd create hypothesis,

33:05

right? Why should we change on your

33:07

website, given what the agents are

33:09

able to build and understanding around

33:11

the industry and the competitors, in

33:14

order to get you a better conversion? And

33:16

then the idea is that they would go

33:18

all the way to implementing those AB tasks

33:20

and measure it so that you can then

33:22

choose. So it's basically automating the whole

33:25

kind of like AB testing kind of

33:27

like process from SEO to copy and

33:29

everything, but going this one step beyond

33:32

kind of like understanding the industry and

33:34

the competitors and everything. And again, something

33:36

that you could do yourself, but that

33:39

would take quite a lot of time.

33:41

So it's basically. giving you recommendations on what

33:43

AB test to run by ingesting that data

33:45

and actually doing the AB test themselves. It's

33:47

actually going ahead and just so you just

33:49

basically come in and updating the code. Yeah.

33:52

Yeah. Explain to what the agent does. Do

33:54

you have to start off by giving it

33:56

data? Like what's the human doing and what's

33:58

the agent doing? So the input. there was

34:00

this is my website this is my description

34:02

of my industry this is what I'm trying

34:04

to optimize this is the human input okay

34:07

and then the agents would go around and

34:09

do everything behind the scenes and they would

34:11

come back with all this like AB hypothesis

34:13

for testing that's pretty awesome and then what

34:16

the company that we were working with was

34:18

actually doing is doing a product around this

34:20

right so you would be able to see

34:22

how the hypothesis and you would say yes

34:24

let's do this one this one this one

34:27

and then they would run for a few

34:29

days and you could implement it but that

34:31

was like there was not the agents anymore

34:33

that was the proper product that they find

34:35

you in the agent or use rag or

34:38

anything to build the hypothesis like a you

34:40

know or was it just the public information

34:42

the the agents knows everything about the internet

34:44

public information yeah okay yeah I think there

34:47

was something about they're doing with the images

34:49

that was kind like more proprietary to them

34:51

like the parsing the images where they're doing

34:53

something fancy with there's not only this not

34:55

only the screenshot but there's also the HDML

34:58

and then they would kind of like create

35:00

a better understanding of like the web page

35:02

by doing that but no that was basically

35:04

a lot of kind of like the open

35:06

models that you have now. Very cool this

35:09

has been a great conversation I wonder if

35:11

where we should end is okay we started

35:13

with you know explaining the agent how to

35:15

get started How they're getting adopted, but really

35:18

like framing it all and like we do

35:20

believe this is years agent You have some

35:22

like really great stats to show that this

35:24

is really happening, right? This is not all

35:26

hype, but what do you think is hype

35:29

in terms of what you hear in how

35:31

agents are spoken about? What is the mismatch

35:33

you see today and where the technology is

35:35

today and where expectations may be for how

35:37

people want to use agents like where do

35:40

you see the biggest mismatch when you speak

35:42

to people? believe especially this year. I don't

35:44

think it's going to be kind of like

35:46

once and done. Like we're seeing this with

35:48

operator, right? And like now there's a bunch

35:51

of tweets where like people said it's do

35:53

things and kind of like fails me right

35:55

through and you got to step in and

35:57

take over and all that. So I think

36:00

this is not the years where we're going

36:02

to have like completely end-to-end, especially on the

36:04

more high precision kind of processes, kind of

36:06

like, oh, agents are doing everything. I think

36:08

that will be kind of like a step-by-step.

36:11

The other thing is implementing on especially the

36:13

more complex use case is going to be

36:15

a way like a... bigger left than a

36:17

lot of people would live. I think there's

36:19

a lot of glue in these companies nowadays

36:22

and in the soft person and how they

36:24

connect to each other that in order for

36:26

agents you can like be able to navigate

36:28

those paths. You're going to need to have

36:31

like clear code and clear instructions on how

36:33

to do it. And I think like there's

36:35

a reason why everyone is doing the browser

36:37

set of things first, right? That is easier.

36:39

Like there's a common interface. But there's a

36:42

lot of companies out there. They have software.

36:44

There's not even online. It's desktop apps. How

36:46

you handle that. So I think there's going

36:48

to be a lot of more challenges in

36:50

there. So I think this year is going

36:53

to be definitely where we're going to see

36:55

a lot of agents going into production. But

36:57

it's very much early days still. This is

36:59

not the years like where agents are taking

37:02

over the like a workforce. That's not it.

37:04

Right. create and hype still those YouTubeers as

37:06

we realize as YouTubeers ourselves. Everything has to

37:08

be like a dramatic headline. That's the way

37:10

it works. Joe, I think this was an

37:13

incredible run through of agents in a way

37:15

that will make it really easy for people

37:17

to actually understand what's the reality is and

37:19

where they can go get started and obviously

37:21

we would highly recommend they go to a

37:24

platform like crew and build some agents and

37:26

play with the technology. And you have a

37:28

bunch of like cool templates actually that make

37:30

it super easy to start to get inspiration.

37:33

One of the things that people actually struggle

37:35

with, which is why I wanted to really

37:37

dig into like how you would suggest someone

37:39

starts with a use case, is people just

37:41

struggle with like where do I even get

37:44

started? I figure template gallery is a really

37:46

great way to find inspiration for your role.

37:48

So I appreciate you coming on and taking

37:50

us through that. The explanation of

37:52

of No worries. Thank

37:55

you so much for

37:57

having me. I had

37:59

a blast. Thank you

38:01

so much much catch I catch

38:03

my Thank you, Joe. you,

38:06

Joe. your help. your help.